A goal that runs for weeks is a different thing to govern
One line in Salesforce's September 11 announcement matters more than the seven named agents around it. Salesforce says it has built a new long-horizon runtime so agents can pursue goals across days and weeks instead of single tasks. The worked example is Hunter, the outbound sales agent, which a seller can ask to rescue their at-risk deals before the end of the quarter. Hunter is in pilot now, with general availability due in November 2026.
A service conversation with Casey, the help agent, ends inside minutes, and the transcript is the whole record. A goal that runs until quarter end will outlast the data it started with and probably its day-one rules. The post says a lot about how such an agent decides and much less about how you find a run, who owns it, and who can stop it.
The controls Salesforce names sit close to the agent's reasoning. Agent Script, an open-source language for agent behavior, combines AI reasoning with deterministic rules. Agent Optimizer, due in October 2026, reads session traces. Hunter itself is described as working out which tasks to complete and which guardrails decide when it acts alone and when seller approval is required.
Each of those governs a single decision. None of them, in the post, answers what an operations lead asks in week three: how many goals are running right now, who started each one, and which are waiting on a person. Session traces will tell you afterwards. The mid-run view is the gap.
Give the run a record you can open
Our read is that a long-horizon goal has to be a record in its own right, with the status discipline you would demand of a case or an opportunity. Who started it, with what goal text, under whose permissions. Current state, last action, and why it is waiting if it is waiting. A planned end date, because before the end of the quarter has to resolve to a day on which the agent stops by itself.
The post credits Hunter with building 60 percent of Perk's sales pipeline, so somebody there has already needed that view. We made the same argument about Power Platform in the case for a durable process record. If the run only exists inside the runtime, the first time finance asks why an account got a discount in week four, you will be reading traces.
Whose permissions apply in week three
The Hunter example is anchored to a seller. Over six weeks that seller goes on leave or hands the territory to a colleague. The post does not say whether the agent's permissions are checked once at the start or before each action, and that is the first thing we'd ask. A run that keeps the starting seller's access for its whole life is an access exception nobody signed off.
The same applies to the rules Agent Script encodes. If the discount policy changes on the first of the month, does a run started on the twenty-fifth pick up the new rule or finish under the old one? We raised a version of this about lifecycle ownership for longer-running agents during the Winter '27 preview.
An approval that waits three days is a different approval
In a chat, approval is a prompt the user answers now. In a run that spans weeks, a seller approval request may sit from Friday to Tuesday. The agent could wait, carry on with the parts that need no sign-off, or expire the request and plan again. Any of those can be defended. Finding out which one you got from a customer complaint cannot.
Then there is the stale approval. A seller approves an outreach sequence on Tuesday, and on Thursday the account raises a support case and asks not to be contacted. Whether that approval still holds depends on whether the runtime rechecks conditions before acting or only records that consent was given. Amani Reed's piece on designing the human review step argues the reviewer should see what the agent will do next. For a long run, that has to include when.
Someone has to stop one run without stopping them all
Dynamic steering, as the post describes it, responds to one user's direction, which means a seller nudging their own agent. An admin who has to freeze every Hunter run on one product line after a pricing error needs more, and so does whoever handles a legal hold on one account. Both need a pause that keeps the record intact and a resume that plans again from current data, for one run, one agent type, or the whole org. We made the case for a pause button in June. This runtime is where Salesforce customers will feel its absence first.
What to ask before November
With Hunter's general availability two months out, we'd take four questions to the account team. Can an admin list every running goal with its owner and planned end date? Are permissions and business rules evaluated at each action or fixed when the run starts? What does the agent do while an approval sits unanswered, and does it expire? Can one run be paused and resumed with its history intact, and can a whole agent type be paused at once?
The companion piece on defining the handoff when an agent works across functions covers the multi-agent half. The two meet at the record. An agent that passes work to Marshall in week two needs somewhere to write that down that a person can find in week five. Put the run record on the Agentforce governance list this month, before the first seller asks Hunter to rescue a quarter.
